## ---- stress_test ----------------------------------------------------

test_that("stress_test gives maximal robustness for zero perturbation", {
  baseline <- c(1, 2, 3, 4, 5)
  out <- stress_test(baseline, list(identical = baseline))
  expect_equal(out$robustness_index, 1)
  expect_equal(out$by_perturbation$identical$mean_abs_change, 0)
  expect_equal(out$by_perturbation$identical$rmse, 0)
})

test_that("stress_test robustness_index decreases with larger perturbation", {
  set.seed(20)
  baseline <- rnorm(100, mean = 5, sd = 1)
  small <- baseline + rnorm(100, mean = 0.1, sd = 0.05)
  large <- baseline + rnorm(100, mean = 1.0, sd = 0.05)
  out_small <- stress_test(baseline, list(p = small))
  out_large <- stress_test(baseline, list(p = large))
  expect_gt(out_small$robustness_index, out_large$robustness_index)
})

test_that("stress_test errors on mismatched perturbation length", {
  expect_error(stress_test(1:5, list(bad = 1:4)))
})

## ---- prompt_sensitivity --------------------------------------------

test_that("identical paraphrase outputs imply zero within-prompt variance", {
  scores <- list(
    version1 = c(1, 2, 3, 4),
    version2 = c(1, 2, 3, 4),
    version3 = c(1, 2, 3, 4)
  )
  out <- prompt_sensitivity(scores)
  expect_equal(out$within_prompt_variance, 0)
  expect_true(is.na(out$sensitivity_ratio) || out$sensitivity_ratio == 0)
})

test_that("prompt_sensitivity flags high within-prompt variance", {
  set.seed(21)
  # same underlying prompts but wildly different across paraphrases
  scores <- list(
    v1 = c(1, 2, 3, 4),
    v2 = c(4, 3, 2, 1),
    v3 = c(2, 4, 1, 3)
  )
  out <- prompt_sensitivity(scores)
  expect_gt(out$within_prompt_variance, 0)
})

test_that("prompt_sensitivity errors on unequal-length paraphrase groups", {
  expect_error(prompt_sensitivity(list(a = 1:4, b = 1:3)), "equal length")
})

test_that("prompt_sensitivity warns and returns NA ratio when between variance is zero", {
  # every underlying prompt has the same mean across paraphrases -> zero between var
  scores <- list(v1 = c(1, 1, 1), v2 = c(2, 2, 2))
  # row means are all 1.5 -> between-prompt variance zero
  expect_warning(out <- prompt_sensitivity(scores), "zero")
  expect_true(is.na(out$sensitivity_ratio))
})

## ---- adversarial_stability -----------------------------------------

test_that("adversarial_stability reports zero flips when outputs are unchanged", {
  orig <- c("A", "B", "A", "C")
  adv  <- c("A", "B", "A", "C")
  out <- adversarial_stability(orig, adv)
  expect_equal(out$flip_rate, 0)
  expect_equal(out$n, 4)
})

test_that("adversarial_stability reports full flip rate when all change", {
  orig <- c("A", "B", "A", "C")
  adv  <- c("B", "A", "C", "A")
  out <- adversarial_stability(orig, adv)
  expect_equal(out$flip_rate, 1)
})

test_that("adversarial_stability accepts a custom flip function", {
  orig <- c(1.0, 2.0, 3.0)
  adv  <- c(1.05, 2.5, 3.0)
  # flip only if difference exceeds 0.4
  out <- adversarial_stability(orig, adv, flip_fn = function(a, b) abs(a - b) > 0.4)
  expect_equal(out$flip_rate, 1/3)
})

test_that("adversarial_stability errors on mismatched lengths", {
  expect_error(adversarial_stability(1:3, 1:2))
})

## ---- ai_robustness (wrapper) ---------------------------------------

test_that("ai_robustness wraps stress_test with the robustness class", {
  baseline <- c(1, 2, 3, 4, 5)
  out <- ai_robustness(baseline, list(p = baseline + 0.1))
  expect_s3_class(out, "aiEvalR_robustness")
  expect_true("robustness_index" %in% names(out))
})